Face Detection Algorithm Based On Lightweight Convolutional Neural Network

Face detection is a research hotspot in the field of computer vision and plays an increasingly important role in daily work and life.For example,intelligent security,robot vision,entertainment beauty and other fields are inseparable from face detection technology.In practical applications,the face detection tasks need to be performed on the hardware platform or embedded platform with limited computing resources.Although traditional face detection algorithms have improved detection accuracy,they ignore the real-time detection,and have the problems of large network model size and low operation efficiency,which can not meet the requirements of transplanting to low-power hardware platform or embedded platform.therefore,research on high precision,light weight and real-time face detection algorithm is very important.This paper proposes a face detection algorithm based on SSD framework,using lightweight convolutional neural network as backbone network,to ensure enough feature extracting and can greatly reduce the network complexity,effectively improve the operation efficiency.Considering the particularity of face shape and proportion,in this paper,the anchor box of number,size and proportion of adjustment,make the face more suitable anchor box matching,thus improve the performance of face detection network.The algorithm in this paper is compared with other algorithms to test the face data set in the monitoring scene on the low-cost Ge Force GTX1080 hardware platform and the embedded Jetson TX1 platform.The results show that:Lightweight face detection algorithm is proposed in this paper,on the premise of guarantee a certain accuracy,the model size is only 2.7 MB,compared to the S~3FD algorithm,SSD algorithm,Mobile Net SSD algorithm model size reduced 87.1MB,92.3 MB,19.5 MB.The detection speed in embedded Jetson TX1 platform reaches 25FPS,and its speed is about 8 times of S~3FD algorithm and SSD algorithm,about 4 times of Mobile Net SSD algorithm,meet the requirements of real-time detection.On the basis of the above improved scheme,the feature fusion module is introduced to fuse the semantic information of shallow and deep features,so as to further improve the accuracy of face detection algorithm.The experimental results show that the face detection algorithm based on feature fusion proposed in this paper has a precision and recall rate of97.88%and 98.28%respectively for the self-built face data set,and is superior to Mobile Net SSD algorithm in detection accuracy,model size and detection speed.The two face detection algorithms proposed in this paper,while ensuring certain accuracy,greatly reduce the complexity of the network and improve the speed of face detection,can run efficiently on low-cost hardware platform and embedded platform.